app.py

from unsloth import FastLanguageModel
from peft import PeftModel

# Load the base model with FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Llama-3.2-3B-Instruct",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True
)

base_model_name = "unsloth/Llama-3.2-3B-Instruct"
adapter_path = "jaspersands/model"  # Path to LoRA adapter on Hugging Face

model = PeftModel.from_pretrained(model, adapter_path)

# Code for processing a query
import pandas as pd
from unsloth.chat_templates import get_chat_template
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from sentence_transformers import SentenceTransformer, util
import nltk

# Ensure you have NLTK stopwords downloaded
nltk.download("stopwords")
from nltk.corpus import stopwords

# Step 1: Load the CSV file
file_path = 'Clean Missouri Data.csv'
df = pd.read_csv(file_path, encoding='MacRoman')

# Step 2: Define a function to search relevant policies based on the user's query using cosine similarity
def search_relevant_policies(query, df, top_n=10, max_chars = 40000):
    # Convert policies into a TF-IDF matrix
    tfidf = TfidfVectorizer(stop_words='english')
    tfidf_matrix = tfidf.fit_transform(df['Content'])

    # Get the query as a TF-IDF vector
    query_vector = tfidf.transform([query])

    # Calculate cosine similarity between query and policies
    cosine_sim = cosine_similarity(query_vector, tfidf_matrix).flatten()

    # Get the top N relevant policies
    top_indices = cosine_sim.argsort()[-top_n:][::-1]
    relevant_policies = df.iloc[top_indices]

    top_indices = cosine_sim.argsort()[-top_n:][::-1]
    relevant_policies = df.iloc[top_indices].copy()

    # Ensure total text is capped at max_chars
    char_count = 0
    valid_indices = []

    for idx, row in relevant_policies.iterrows():
        content_length = len(row["Content"])

        # If adding this content exceeds max_chars, stop adding any further policies
        if char_count + content_length > max_chars:
            break

        # Otherwise, keep this policy
        char_count += content_length
        valid_indices.append(idx)

    # Filter the dataframe to only include valid rows
    truncated_policies = relevant_policies.loc[valid_indices]

    return truncated_policies


def get_content_after_query(response_text, query):
    # Find the position of the query within the response text
    query_position = response_text.lower().find(query.lower())
    if query_position != -1:
        # Return the content after the query position
        res = response_text[query_position + len(query):].strip()
        return res[11:]
    else:
        # If the query is not found, return the full response text as a fallback
        return response_text.strip()


def process_query(query,tokenizer):

    relevant_policies = search_relevant_policies(query, df)

    # Step 5: Combine the relevant policies with the user's query for the model
    formatted_policies = []
    for index, row in relevant_policies.iterrows():
        # formatted_policy = f"Title: {row['Title']}\nTerritory: {row['Territory']}\nType: {row['Type']}\nYear: {row['Year']}\nCategory: {row['Category']}\nFrom: {row['From']}\nTo: {row['To']}\nContent: {row['Content']}\nLink: {row['Link to Content']}\n"
        # formatted_policies.append(formatted_policy)
        formatted_policies.append(row['Content'])
    relevant_policy_text = "\n\n".join(formatted_policies)

    # Messages with relevant policies for the model
    messages_with_relevant_policies = [
        {"role": "system", "content": relevant_policy_text},
        {"role": "user", "content": query},
    ]

    # Step 6: Apply chat template and tokenize
    tokenizer = get_chat_template(
        tokenizer,
        chat_template="llama-3.1",
    )
    inputs = tokenizer.apply_chat_template(
        messages_with_relevant_policies,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to("cuda")

    FastLanguageModel.for_inference(model)
    outputs = model.generate(input_ids=inputs, max_new_tokens=512, use_cache=True, temperature=0.7, min_p=0.1)

    # Step 7: Decode the output
    generated_response = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    response = get_content_after_query(generated_response, query)

    # Step 8: Rank the top 10 policies using SBERT for the final link
    # Load SBERT model
    model_sbert = SentenceTransformer('all-MiniLM-L6-v2')  # You can choose another SBERT model if desired

    # Encode the generated response using SBERT
    response_embedding = model_sbert.encode(generated_response, convert_to_tensor=True)

    # Encode each policy in the top 10 list
    policy_embeddings = model_sbert.encode(relevant_policies['Content'].tolist(), convert_to_tensor=True)

    # Calculate cosine similarities between the generated response and each policy embedding
    cosine_similarities = util.cos_sim(response_embedding, policy_embeddings).flatten()

    # Identify the policy with the highest SBERT cosine similarity score
    most_relevant_index = cosine_similarities.argmax().item()
    most_relevant_link = relevant_policies.iloc[most_relevant_index]['Link to Content']

    # Print the link to the most relevant source
    return {
        "response": response,
        "most_relevant_link": most_relevant_link
    }


    # Load Google Sheets to store results
import json
from google.oauth2.service_account import Credentials
import gspread
import pandas as pd

# Load the service account JSON
json_file_path = "fostercare-449201-75a303a8c238.json"  # Load the credentials for the service account
with open(json_file_path, 'r') as file:
    service_account_data = json.load(file)

# Authenticate using the loaded service account data
scopes = ["https://www.googleapis.com/auth/spreadsheets", "https://www.googleapis.com/auth/drive"]
creds = Credentials.from_service_account_info(service_account_data, scopes=scopes)
client = gspread.authorize(creds)

# Open the shared Google Sheet by name
spreadsheet = client.open("Foster Care RA Responses").sheet1

# Link to Google Sheet
# https://docs.google.com/spreadsheets/d/15iEcxmTgkgfcxzDGnq3i_nP1hiAXgb3RplHgqAMEyHA/edit?usp=sharing


# Code to set up Gradio GUI
import gradio as gr

def greet(query):
    result_1 = process_query(query, tokenizer)
    content_after_query_1 = result_1["response"]

    result_2 = process_query(query, tokenizer)
    content_after_query_2 = result_2["response"]

    return [content_after_query_1, content_after_query_2]

def choose_preference(name, output1, output2, preference, query):
    if not name:
      return "Please enter your name before submitting."

    if preference == "Output 1":
        new_row = [query, output1, output2, name]
        spreadsheet.append_row(new_row)
        return f"You preferred: Output 1 - {output1}"
    elif preference == "Output 2":
        new_row = [query, output2, output1, name]
        spreadsheet.append_row(new_row)
        return f"You preferred: Output 2 - {output2}"
    else:
        return "No preference selected."

# Define the interface
with gr.Blocks() as demo:
    # Name input
    name_input = gr.Textbox(label="Enter your name")

    # Input for query
    query_input = gr.Textbox(label="Enter your query")

    # Outputs
    output_1 = gr.Textbox(label="Output 1", interactive=False)
    output_2 = gr.Textbox(label="Output 2", interactive=False)

    # Preference selection
    preference = gr.Radio(["Output 1", "Output 2"], label="Choose your preferred output")
    preference_result = gr.Textbox(label="Your Preference", interactive=False)

    # Buttons
    generate_button = gr.Button("Generate Outputs")
    submit_button = gr.Button("Submit Preference")

    # Link actions to buttons
    generate_button.click(greet, inputs=query_input, outputs=[output_1, output_2])
    submit_button.click(choose_preference, inputs=[name_input, output_1, output_2, preference, query_input], outputs=preference_result)

demo.launch()